The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to launch and interact with terminal applications using tmux. The scripts/tmux-real-user-log.sh script executes arbitrary commands passed as arguments to the start command via tmux new-session.
[DATA_EXFILTRATION]: The skill utilizes tmux capture-pane to record the entire terminal state to plain-text log files. The instructions specifically suggest this for /auth and other interactive workflows. This may result in sensitive information, such as passwords or authentication tokens, being stored unencrypted in the tmp/ directory. The instructions explicitly advise against sanitizing these logs unless requested.
[DYNAMIC_EXECUTION]: The skill promotes the use of eval "$(bash ...)" to import variables from the helper script's output into the current shell session. This execution of dynamically generated shell code poses a risk if the script output is tampered with or contains unexpected content.
[INDIRECT_PROMPT_INJECTION]: The agent is instructed to poll and read the content of the tmux pane to determine the test status. If the application under test displays untrusted or malicious content, it could potentially inject instructions into the agent's context.
Ingestion points: tmux capture-pane output saved to tmux-readable-full.log and current-pane.txt in SKILL.md and scripts/tmux-real-user-log.sh.
Boundary markers: Absent; terminal snapshots are appended directly to the log without delimiters or isolation.
Capability inventory: Shell command execution (via tmux), file system read/write, and environment variable modification.
Sanitization: None; the skill captures and processes raw rendered terminal output.

tmux-real-user-testing